From Probability, For the Enthusiastic Beginner (Draft version, March 2016) David Morin,

Transcription

1 Chapter 4 Distributions From Probability, For the Enthusiastic Beginner (Draft version, March 2016) David Morin, morin@physics.harvard.edu At the beginning of Section 3.1, we introduced the concepts of random variables and probability distributions. A random variable is a variable that can take on certain numerical values with certain probabilities. The collection of these probabilities is called the probability distribution for the random variable. A probability distribution specifies how the total probability (which is always 1) is distributed among the various possible outcomes. In this chapter, we will discuss probability distributions in detail. In Section 4.1 we warm up with some examples of discrete distributions, and then in Section 4.2 we discuss continuous distributions. These involve the probability density, which is the main new concept in this chapter. It takes some getting used to, but we ll have plenty of practice with it. In Sections we derive and discuss a number of the more common and important distributions. They are, respectively, the uniform, Bernoulli, binomial, exponential, Poisson, and Gaussian (or normal) distributions. Parts of this chapter are a bit mathematical, but there s no way around this if we want to do things properly. However, we ve relegated some of the more technical issues to Appendices B and C. If you want to skip those and just accept the results that we derive there, that s fine. But you are strongly encouraged to at least take a look at Appendix B, where we derive many properties of the number e, which is the most important number in probability and statistics. 4.1 Discrete distributions In this section we ll give a few simple examples of discrete distributions. To start off, consider the results from Example 3 in Section 2.3.4, where we calculated the probabilities of obtaining the various possible numbers of Heads in five coin flips. We found: P(0) = 1 32, P(1) = 5, P(2) = 32, P(3) = 10 32, P(4) = 5 32, P(5) = (4.1) 182

2 4.1. Discrete distributions 183 These probabilities add up to 1, as they should. Fig. 4.1 shows a plot of P(n) versus n. The random variable here is the number of Heads, and it can take on the values of 0 through 5, with the above probabilities. P(n) 10/32 5/ n Figure 4.1: The probability distribution for the number of Heads in five coin flips. As we ve done in Fig. 4.1, the convention is to plot the random variable on the horizontal axis and the probability on the vertical axis. The collective information, given either visually in Fig. 4.1 or explicitly in Eq. (4.1), is the probability distribution. A probability distribution simply tells you what all the probabilities are for the values that the random variable can take. Note that P(n) in the present example is nonzero only if n takes on one of the discrete values, 0, 1, 2, 3, 4, or 5. It s a silly question to ask for the probability of getting 4.27 Heads, because n must of course be an integer. The probability of getting 4.27 Heads is trivially zero. Hence the word discrete in the title of this section. Another simple example of a discrete probability distribution is the one for the six possible outcomes of the roll of one die. The random variable in this setup is the number on the top face of the die. If the die is fair, then all six numbers have equal probabilities, so the probability for each is 1/6, as shown in Fig P(n) 1/ n Figure 4.2: The probability distribution for the roll of one die. What if the die isn t fair? For example, what if we make the 1 face heavier than the others by embedding a small piece of lead in the center of that face, just below the surface? The die is then more likely to land with the 1 face pointing down. The 6 face is opposite the 1, so the die is more likely to land with the 6 pointing up. Fig. 4.2 will therefore be modified by raising the 6 dot and lowering

3 184 Chapter 4. Distributions the other five dots; the sum of the probabilities must still be 1, of course. P 2 through P 5 are all equal, by symmetry. The exact values of all the probabilities depend in a complicated way on how the mass of the lead weight compares with the mass of the die, and also on the nature of both the die and the table on which the die is rolled (how much friction, how bouncy, etc.). As mentioned at the beginning of Section 3.1, a random variable is assumed to take on numerical values, by definition. So the outcomes of Heads and Tails for a single coin flip technically aren t random variables. But it still makes sense to plot the probabilities as shown in Fig. 4.3, even though the outcomes on the horizontal axis aren t associated with a random variable. Of course, if we define a random variable to be the number of Heads, then the Heads in the figure turns into a 1, and the Tails turns into a 0. In most situations, however, the outcomes take on numerical values right from the start, so we can officially label them as random variables. But even if they don t, we ll often take the liberty of still referring to the thing being plotted on the horizontal axis of a probability distribution as a random variable. P(face) 1/2 Tails Heads face Figure 4.3: The probability distribution for a single coin flip. 4.2 Continuous distributions Motivation Probability distributions are fairly straightforward when the random variable is discrete. You just list (or plot) the probabilities for each of the possible values of the random variable. These probabilities will always add up to 1. However, not everything comes in discrete quantities. For example, the temperature outside your house takes on a continuous set of values, as does the amount of water in a glass. (We ll ignore the atomic nature of matter!) In finding the probability distribution for a continuous random variable, you might think that the procedure should be exactly the same as in the discrete case. That is, if our random variable is the temperature at a particular location at noon tomorrow, then you might think that you simply have to answer questions of the form: What is the probability that the temperature at noon tomorrow will be 70 Fahrenheit?

4 4.2. Continuous distributions 185 Unfortunately, there is something wrong with this question, because it is too easy to answer. The answer is that the probability is zero, because there is simply no chance that the temperature at a specific time (and a specific location) will be exactly 70. If it s 70.1, that s not good enough. And neither is 70.01, nor even Basically, since the temperature takes on a continuous set of values (and hence an infinite number of possible values), the probability of a specific value occurring is 1/, which is zero.1 However, even though the above question ( What is the probability that the temperature at noon tomorrow will be 70? ) is a poor one, that doesn t mean we should throw in the towel and conclude that probability distributions don t exist for continuous random variables. They do in fact exist, because there are some useful questions we can ask. These useful questions take the general form of: What is the probability that the temperature at a particular location at noon tomorrow lies somewhere between 69 and 71? This question has a nontrivial answer, in the sense that it isn t automatically zero. And depending on what the forecast is for tomorrow, the answer might be something like 20%. We can also ask: What is the probability that the temperature at noon lies somewhere between 69.5 and 70.5? The answer to this question is smaller than the answer to the previous one, because it involves a range of only one degree instead of two degrees. If we assume that inside the range of 69 to 71 the temperature is equally likely to be found anywhere (which is a reasonable approximation although undoubtedly not exactly correct), and if the previous answer was 20%, then the present answer is (roughly) 10%, because the range is half the size. The point here is that the smaller the range, the smaller the chance that the temperature lies in that range. Conversely, the larger the range, the larger the chance that the temperature lies in that range. Taken to an extreme, if we ask for the probability that the temperature at noon lies somewhere between 100 and 200, then the answer is exactly equal to 1 (ignoring liquid nitrogen spills, forest fires, and such things!). In addition to depending on the size of the range, the probability also of course depends on where the range is located on the temperature scale. For example, the probability that the temperature at noon lies somewhere between 69 and 71 is undoubtedly different from the probability that it lies somewhere between 11 and 13. Both ranges have a span of two degrees, but if the given day happens to be in late summer, the temperature is much more likely to be around 70 than to be sub-freezing (let s assume we re in, say, Boston). To actually figure out the probabilities, many different pieces of data would have to be considered. In the present temperature example, the data would be of the meteorological type. But if we were interested in the probability that a random person is between 69 and 71 inches tall, then we d need to consider a whole different set of data. The lesson to take away from all this is that if we re looking at a random variable that can take on a continuous set of values, the probability that this random variable falls into a given range depends on three things. It depends on: 1Of course, if you re using a digital thermometer that measures the temperature to the nearest tenth of a degree, then it does make sense to ask for the probability that the thermometer reads, say, 70.0 degrees. This probability is generally nonzero. This is due to the fact that the reading on the digital thermometer is a discrete random variable, whereas the actual temperature is a continuous random variable.

5 186 Chapter 4. Distributions 1. the location of the range, 2. the size of the range, 3. the specifics of the situation we re dealing with. The third of these is what determines the probability density, which is a function whose argument is the location of the range. We ll now discuss probability densities Probability density Consider the plot in Fig. 4.4, which gives a hypothetical probability distribution for the temperature example we ve been discussing. This plot shows the probability distribution on the vertical axis, as a function of the temperature T (the random variable) on the horizontal axis. We have chosen to measure the temperature in Fahrenheit. We re denoting the probability distribution by2 ρ(t) instead of P(T), to distinguish it from the type of probability distribution we ve been talking about for discrete variables. The reason for this new notation is that ρ(t) is a probability density and not an actual probability. We ll talk about this below. When writing the functional form of a probability distribution, we ll denote probability densities with lowercase letters, like the ρ in ρ(t) or the f in f (x). And we ll denote actual probabilities with uppercase letters, like the P in P(n). 0.1 ρ(t) T Figure 4.4: A hypothetical probability distribution for the temperature. We haven t yet said exactly what we mean by ρ(t). But in any case, it s clear from Fig. 4.4 that the temperature is more likely to be near 70 than near 60. The following definition of ρ(t) allows us to be precise about what we mean by this. 2As mentioned at the beginning of Section 3.1, a random variable is usually denoted with an uppercase letter, while the actual values are denoted with lowercase letters. So we should technically be writing ρ(t) here. But since an uppercase T is the accepted notation for temperature, we ll use T for the actual value.

6 4.2. Continuous distributions 187 Definition of the probability density function, ρ(t): ρ(t) is the function of T that, when multiplied by a small interval T, gives the probability that the temperature lies between T and T + T. That is, P(temp lies between T and T + T) = ρ(t) T. (4.2) Note that the lefthand side contains an actual probability P, whereas the righthand side contains a probability density, ρ(t). The latter needs to be multiplied by a range of T (or whatever quantity we re dealing with) in order to obtain an actual probability. The above definition is relevant to any continuous random variable, of course, not just temperature. Eq. (4.2) might look a little scary, but a few examples should clear things up. From Fig. 4.4, it looks like ρ(70 ) is about So if we pick T = 1, we find that the probability of the temperature lying between 70 and 71 is about ρ(t) T = (0.07)(1) = 0.07 = 7%. (4.3) If we instead pick a smaller T, say 0.5, we find that the probability of the temperature lying between 70 and 70.5 is about (0.07)(0.5) = 3.5%. And if we pick an even smaller T, say 0.1, we find that the probability of the temperature lying between 70 and 70.1 is about (0.07)(0.1) = 0.7%. Similarly, we can apply Eq. (4.2) to any other value of T. For example, it looks like ρ(60 ) is about So if we pick T = 1, we find that the probability of the temperature lying between 60 and 61 is about (0.02)(1) = 2%. And as above, we can pick other values of T too. Note that, in accordance with Eq. (4.2), we have been using the value of ρ at the lower end of the given temperature interval. That is, when the interval was 70 to 71, we used ρ(70 ) and then multiplied this by T. But couldn t we just as well use the value of ρ at the upper end of the interval? That is, couldn t the righthand side of Eq. (4.2) just as well be ρ(t + T) T? Indeed it could. But as long as T is small, it doesn t matter much which value of ρ we use. They will both give essentially the same answer. See the second remark below. Remember that three inputs are necessary when finding the probability that the temperature lies in a specified range. As we noted at the end of Section 4.2.1, the first input is the value of T we re concerned with, the second is the range T, and the third is the information encapsulated in the probability density function, ρ(t), evaluated at the given value of T. The latter two of these three quantities are the two quantities that are multiplied together on the righthand side of Eq. (4.2). Knowing only one of these isn t enough to give you a probability. To recap, there is a very important difference between the probability distribution for a continuous random variable and that for a discrete random variable. For a continuous variable, the probability distribution consists of a probability density. But for a discrete variable, it consists of actual probabilities. We plot a density for a continuous distribution, because it wouldn t make sense to plot actual probabilities, since they re all zero. This is true because the probability of obtaining exactly a particular value is zero, since there is an infinite number of possible values. Conversely, we plot actual probabilities for a discrete distribution, because it wouldn t make sense to plot a density, since it consists of a collection of infinite

7 188 Chapter 4. Distributions spikes. This is true because on a die roll, for example, there is a 1/6 chance of obtaining a number between, say, and The probability density at the outcome of 5, which from Eq. (4.2) equals the probability divided by the interval length, is then (1/6)/( ), which is huge. And the interval can be made arbitrarily small, which means that the density is arbitrarily large. To sum up, the term probability distribution applies to both continuous and discrete variables, whereas the term probability density applies only to continuous variables. Remarks: 1. ρ(t) is a function of T, so it depends on what units we re using to measure T. We used Fahrenheit above, but what if we instead want to use Celsius? Problem 4.1 addresses this issue (but you will need to read Section first). 2. Note the inclusion of the word small in the definition of the probability density in Eq. (4.2). The reason for this word is that we want ρ(t) to be (roughly) constant over the specified range. If T is small enough, then this is approximately true. If ρ(t) varied greatly over the range of T, then it wouldn t be clear which value of ρ(t) we should multiply by T to obtain the probability. The point is that if T is small enough, then all of the ρ(t) values are roughly the same, so it doesn t matter which one we pick. An alternative definition of the density ρ(t) is P ( temp lies between T ( T)/2 and T + ( T)/2 ) = ρ(t) T. (4.4) The only difference between this definition and the one in Eq. (4.2) is that we re now using the value of ρ(t) at the midpoint of the temperature range, instead of the leftend value we used in Eq. (4.2). Both definitions are equally valid, because they give essentially the same result for ρ(t), provided that T is small. Similarly, we could use the value of ρ(t) at the right end of the temperature range. How small do we need T to be? The answer to this will be evident when we talk about probability in terms of area in Section In short, we need the change in ρ(t) over the span of T to be small compared with the values of ρ(t) in that span. 3. The probability density function involves only (1) the value of T (or whatever) we re concerned with, and (2) the specifics of the situation at hand (meteorological data in the above temperature example, etc.). The density is completely independent of the arbitrary value of T that we choose. This is how things work with any kind of density. For example, consider the mass density of gold. This mass density is a property of the gold itself. More precisely, it is a function of each point in the gold. For pure gold, the density is constant throughout the volume, but we could imagine impurities that would make the mass density be a varying function of position, just as the above probability density is a varying function of temperature. Let s call the mass density ρ(r), where r signifies the possible dependence of ρ on the location of a given point within the volume. (The position of a given point can be described by the vector pointing from the origin to the point. And vectors are generally denoted by boldface letters like r.) Let s call the small volume we re concerned with V. Then the mass in the small volume V is given by the product of the density and the volume, that is, ρ(r) V. This is directly analogous to the fact that the probability in the above temperature example is given by the product of the probability density and the temperature span,

8 4.2. Continuous distributions 189 that is, ρ(t) T. The correspondence among the various quantities is Mass in V around location r Prob that temp lies in T around T ρ(r) ρ(t) V T. (4.5) Probability equals area The graphical interpretation of the product ρ(t) T in Eq. (4.2) is that it is the area of the rectangle shown in Fig This is true because T is the base of the rectangle, and ρ(t) is the height. 0.1 ρ(t) T Figure 4.5: Interpretation of the product ρ(t) T as an area. We have chosen T to be 2 in the figure. With this choice, the area of the rectangle, which equals ρ(70 ) (2 ), gives a reasonably good approximation to the probability that the temperature lies between 70 and 72. But it isn t exact, because ρ(t) isn t constant over the 2 interval. A better approximation to the probability that the temperature lies between 70 and 72 is achieved by splitting the 2 interval into two intervals of 1 each, and then adding up the probabilities of lying in each of these two intervals. These two probabilities are approximately equal to ρ(70 ) (1 ) and ρ(71 ) (1 ), and the two corresponding rectangles are shown in Fig But again, the sum of the areas of these two rectangles is still only an approximate result for the true probability that the temperature lies between 70 and 72, because ρ(t) isn t constant over the 1 intervals either. A better approximation is achieved by splitting the 1 intervals into smaller intervals, and then again into even smaller ones. And so on. When we get to the point of having 100 or 1000 extremely thin rectangles, the sum of their areas will essentially be the area shown in Fig This area is the correct probability that the temperature lies between 70 and 72. So in retrospect, we see that the rectangular area in Fig. 4.5 exceeds the true probability by the area of the tiny triangular-ish region in the upper righthand corner of the rectangle. We therefore arrive at a more precise definition (compared with Eq. (4.2)) of the probability density, ρ(t):

9 190 Chapter 4. Distributions 0.1 ρ(t) T Figure 4.6: Subdividing the area, to produce a better approximation to the probability. 0.1 ρ(t) T Figure 4.7: The area below the curve between 70 and 72 equals the probability that the temperature lies between 70 and 72. Improved definition of the probability density function, ρ(t): ρ(t) is the function of T for which the area under the ρ(t) curve between T and T + T gives the probability that the temperature (or whatever quantity we re dealing with) lies between T and T + T. This is an exact definition, and there is no need for T to be small, as there was in the definition in Eq. (4.2). The difference is that the present definition involves the exact area, whereas Eq. (4.2) involved the area of a rectangle (via simple multiplication by T), which was only an approximation. But technically the only thing we need to add to Eq. (4.2) is the requirement that we take the T 0 limit. That makes the definition rigorous. The total area under any probability density curve must be 1, because this area equals the probability that the temperature (or whatever) takes on some value between and +, and because every possible result is included in the to + range. However, in any realistic case, the density is essentially zero outside a specific finite region. So there is essentially no contribution to the area from the parts

10 4.3. Uniform distribution 191 of the plot outside that region. There is therefore no need to go to ±. The total area under each of the curves in the above figures, including the tails on either side which we haven t bothered to draw, is indeed equal to 1 (at least roughly; the curves were drawn by hand). Given a probability density function f (x), the cumulative distribution function F(x) is defined to be the probability that X takes on a value that is less than or equal to x. That is, F(x) = P(X x). For a continuous distribution, this definition implies that F(x) equals the area under the f (x) curve from up to the given x value. A quick corollary is that the probability P(a < x b) that x lies between two given values a and b is equal to F(b) F(a). For a discrete distribution, the definition F(x) = P(X x) still applies, but we now calculate P(X x) by forming a discrete sum instead of finding an area. Although the cumulative distribution function can be very useful in probability and statistics, we won t use it much in this book. We ll now spend a fair amount of time in Sections discussing some common types of probability distributions. There is technically an infinite number of possible distributions, although only a hundred or so come up frequently enough to have names. And even many of these are rather obscure. A handful, however, come up again and again in a variety of settings, so we ll concentrate on these. They are the uniform, Bernoulli, binomial, exponential, Poisson, and Gaussian (or normal) distributions. 4.3 Uniform distribution We ll start with a very simple continuous probability distribution, one that is uniform over a given interval, and zero otherwise. Such a distribution might look like the one shown in Fig If the distribution extends from x 1 to x 2, then the value of ρ(x) in that region must be 1/(x 2 x 1 ), so that the total area is 1. 1/(x 2 -x 1 ) ρ(x) x 1 x 2 x Figure 4.8: A uniform distribution. This type of distribution could arise, for example, from a setup where a rubber ball bounces around in an empty rectangular room. When it finally comes to rest, we measure its distance x from a particular one of the walls. If you initially throw the ball hard enough, then it s a pretty good approximation to say that x is equally likely to take on any value between 0 and L, where L is the length of the room in the relevant direction. In this setup, the x 1 in Fig. 4.8 equals 0 (so we would need to shift the rectangle to the left), and the x 2 equals L.

11 192 Chapter 4. Distributions The random variable here is X, and the value it takes is denoted by x. So x is what we plot on the horizontal axis. Since we re dealing with a continuous distribution, we plot the probability density (not the probability!) on the vertical axis. If L equals 10 feet, then outside the region 0 < x < 10, the probability density ρ(x) equals zero. Inside this region, the density equals the total probability divided by the total interval, which gives 1 per 10 feet, or equivalently 1/10 per foot. If we want to find the actual probability that the ball ends up between, say, x = 6 and x = 8, then we just multiply ρ(x) by the interval length, which is 2 feet. The result is (1/10 per foot)(2 feet), which equals 2/10 = 1/5. This makes sense, of course, because the 2-foot interval is 1/5 of the total distance. A uniform density is easy to deal with, because the area under a given part of the curve (which equals the probability) is simply a rectangle. And the area of a rectangle is just the base times the height, which is the interval length times the density. This is exactly the product we formed above. When the density isn t uniform, it can be very difficult sometimes to find the area under a given part of the curve. Note that the larger the region of nonzero ρ(x) in a uniform distribution, the smaller the value of ρ(x). This follows from the fact that the total area under the density curve (which is just a straight line segment in this case) must equal 1. So if the base becomes longer, the height must become shorter. 4.4 Bernoulli distribution We ll now consider a very simple discrete distribution, called the Bernoulli distribution. This is the distribution for a process in which only two possible outcomes, 1 and 0, can occur, with probabilities p and 1 p, respectively. (They must add up to 1, of course.) The plot of this probability distribution is shown in Fig It is common to call the outcome of 1 a success and the outcome of 0 a failure. A special case of a Bernoulli distribution is the distribution for a coin toss, where the probabilities for Heads and Tails (which we can assign the values of 1 and 0, respectively) are both equal to 1/2. p 1-p P 0 1 Figure 4.9: A Bernoulli distribution takes on the values 1 and 0 with probabilities p and 1 p. The Bernoulli distribution is the simplest of all distributions, with the exception of the trivial case where only one possible outcome can occur, which therefore has

12 4.5. Binomial distribution 193 a probability of 1. The uniform and Bernoulli distributions are simple enough that there isn t much to say. In contrast, the distributions in the following four sections (binomial, exponential, Poisson, and Gaussian) are a bit more interesting, so we ll have plenty to say about them. 4.5 Binomial distribution The binomial distribution, which is discrete, is an extension of the Bernoulli distribution. The binomial distribution is defined to be the probability distribution for the total number of successes that arise in an arbitrary number of independent and identically distributed Bernoulli processes. An example of a binomial distribution is the probability distribution for the number of Heads in, say, five coin tosses, which we discussed in Section 4.1. We could just as well pick any other number of tosses. In the case of five coin tosses, each coin toss is a Bernoulli process. When we put all five tosses together and look at the total number of successes (Heads), we get a binomial distribution. Let s label the total number of successes as k. In this specific example, there are n = 5 Bernoulli processes, with each one having a p = 1/2 probability of success. The probability distribution P(k) is simply the one we plotted earlier in Fig. 4.1, where we counted the number of Heads. Let s now find the binomial distribution associated with a general number n of independent Bernoulli trials, each with the same probability of success, p. So our goal is to find the value of P(k) for all of the different possible values of the total number of successes, k. The possible values of k range from 0 up to the number of trials, n. To calculate the binomial distribution (for given n and p), we first note that p k is the probability that a specific set of k of the n Bernoulli processes all yield success, because each of the k processes has a p probability of yielding success. We then need the other n k processes to not yield success, because we want exactly k successes. This happens with probability (1 p) n k, because each of the n k processes has a 1 p probability of yielding failure. The probability that a specific set of k processes (and no others) all yield success is therefore p k (1 p) n k. Finally, since there are ( n k) ways to pick a specific set of k processes, we see that the probability that exactly k of the n processes yield success is P(k) = ( ) n p k (1 p) n k (binomial distribution) (4.6) k This is the desired binomial distribution. Note that this distribution depends on two parameters the number n of Bernoulli trials and the probability p of success in each trial. If you want to make these parameters explicit, you can write the Binomial distribution P(k) as B n, p (k). That is, B n, p (k) = But we ll generally just use the simple P(k) notation. ( ) n p k (1 p) n k. (4.7) k

13 194 Chapter 4. Distributions In the special case of a binomial distribution generated from n coin tosses, we have p = 1/2. So Eq. (4.6) gives the probability of obtaining k Heads as P(k) = 1 ( ) n 2 n. (4.8) k To recap: In Eq. (4.6), n is the total number of Bernoulli processes, p is the probability of success in each Bernoulli process, and k is the total number of successes in the n processes. (So k can be anything from 0 to n.) Fig shows the binomial distribution for the cases of n = 30 and p = 1/2 (which arises from 30 coin tosses), and n = 30 and p = 1/6 (which arises from 30 die rolls, with a particular one of the six numbers representing success). P(k) P(k) k k n = 30, p = 1/2 n = 30, p = 1/6 Figure 4.10: Two binomial distributions with n = 30 but different values of p. Example (Equal probabilities): Given n, for what value of p is the probability of zero successes equal to the probability of one success? Solution: In Eq. (4.6) we want P(0) to equal P(1). This gives ( ) ( ) n p 0 (1 p) n 0 n = p 1 (1 p) n = 1 1 (1 p) n = n p (1 p) n 1 = 1 p = np = p = 1 n + 1. (4.9) This p = 1/(n + 1) value is the special value of p for which various competing effects cancel. On one hand, P(1) contains an extra factor of n from the ( n 1 ) coefficient, which arises from the fact that there are n different ways for one success to happen. But on the other hand, P(1) also contains a factor of p, which arises from the fact that one success does happen. The first of these effects makes P(1) larger than P(0), while the second makes it smaller.3 The effects cancel when p = 1/(n + 1). Fig shows the plot for n = 10 and p = 1/11. The p = 1/(n + 1) case is the cutoff between the maximum of P(k) occurring when k is zero or nonzero. If p is larger than 1/(n + 1), as it is in both plots in Fig Another effect is that P(1) is larger because it contains one fewer factor of (1 p). But this effect is minor when p is small, which is the case if n is large, due to the p = 1/(n + 1) form of the answer.

14 4.5. Binomial distribution 195 P(k) k n = 10, p = 1/11 Figure 4.11: P(0) equals P(1) if p = 1/(n + 1). above, then the maximum occurs at a nonzero value of k. That is, the distribution has a bump. On the other hand, if p is smaller than 1/(n + 1), then the maximum occurs at k = 0. That is, the distribution has its peak at k = 0 and falls off from there. Having derived the binomial distribution in Eq. (4.6), there is a simple double check that we can perform on the result. Since the number of successes, k, can take on any integer value from 0 to n, the sum of the P(k) probabilities from k = 0 to k = n must equal 1. The P(k) expression in Eq. (4.6) does indeed satisfy this requirement, due to the binomial expansion, which tells us that ( ) n n ( ) n p + (1 p) = p k (1 p) n k. (4.10) k k=0 This is just Eq. (1.21) from Section 1.8.3, with a = p and b = 1 p. The lefthand side of Eq. (4.10) is simply 1 n = 1. And each term in the sum on the righthand side is a P(k) term from Eq. (4.6). So Eq. (4.10) becomes 1 = n P(k), (4.11) k=0 as we wanted to show. You are encouraged to verify this result for the probabilities in, say, the left plot in Fig Feel free to make rough estimates of the probabilities when reading them off the plot. You will find that the sum is indeed 1, up to the rough estimates you make. The task of Problem 4.4 is to use Eq. (3.4) to explicitly demonstrate that the expectation value of the binomial distribution in Eq. (4.6) equals pn. In other words, if our binomial distribution is derived from n Bernoulli trials, each having a probability p of success, then we should expect a total of pn successes (on average, if we do a large number of sets of n trials). This must be true, of course, because a fraction p of the n trials yield success, on average, by the definition of p for the given Bernoulli process.

15 196 Chapter 4. Distributions Remark: We should emphasize what is meant by a probability distribution. Let s say that you want to experimentally verify that the left plot in Fig is the correct probability distribution for the total number of Heads that show up in 30 coin flips. You of course can t do this by flipping a coin just once. And you can t even do it by flipping a coin 30 times, because all you ll get from that is just one number for the total number of Heads. For example, you might obtain 17 Heads. In order to experimentally verify the distribution, you need to perform a large number of sets of 30 coin flips, and you need to record the total number of Heads you get in each 30-flip set. The result will be a long string of numbers such as 13, 16, 15, 16, 18, 14, 11, 17,.... If you then calculate the fractions of the time that each number appears, these fractions should (roughly) agree with the probabilities shown in Fig The longer the string of numbers, the better the agreement, in general. The main point here is that the distribution does t say much about one particular set of 30 flips. Rather, it says what the expected distribution of outcomes is for a large number of sets of 30 flips. 4.6 Exponential distribution In Sections we ll look at three probability distributions (exponential, Poisson, and Gaussian) that are a bit more involved than the three we ve just discussed (uniform, Bernoulli, and binomial). We ll start with the exponential distribution, which takes the general form, ρ(t) = Ae bt, (4.12) where A and b are quantities that depend on the specific situation at hand. We will find below in Eq. (4.26) that these quantities must be related in a certain way in order for the total probability to be 1. The parameter t corresponds to whatever the random variable is. The exponential distribution is a continuous one, so ρ(t) is a probability density. The most common type of situation where this distribution arises is the following. Consider a repeating event that happens completely randomly in time. By completely randomly we mean that there is a uniform probability that the event happens at any given instant (or more precisely, in any small time interval of a given length), independent of what has already happened. That is, the process has no memory. The exponential distribution that we ll eventually arrive at (after a lot of work!) in Eq. (4.26) gives the probability distribution for the waiting time until the next event occurs. Since the time t is a continuous quantity, we ll need to develop some formalism to analyze the distribution. To ease into it, let s start with the slightly easier case where time is assumed to be discrete Discrete case Consider a process where we roll a hypothetical 10-sided die once every second. So time is discretized into 1-second intervals. It s actually not necessary to introduce time here at all. We could simply talk about the number of iterations of the process. But it s easier to talk about things like the waiting time than the number of iterations you need to wait for. So for convenience, we ll discuss things in the context of time. If the die shows a 1, we ll consider that a success. The other nine numbers represent failure. There are two reasonable questions we can ask: What is the average

16 4.6. Exponential distribution 197 waiting time (that is, the expectation value of the waiting time) between successes? And what is the probability distribution of the waiting times between successes? Average waiting time It is fairly easy to determine the average waiting time. There are 10 possible numbers on the die, so on average we can expect 1/10 of them to be 1 s. If we run the process for a long time, say, an hour (which consists of 3600 seconds), then we can expect about s. The average waiting time between successes is therefore (3600 seconds)/360 = 10 seconds. More generally, if the probability of success in each trial is p, then the average waiting time is 1/p (assuming that the trials happen at 1-second intervals). This can be seen by the same reasoning as above. If we perform n trials of the process, then pn of them will yield success, on average. The average waiting time between successes is the total time (n) divided by the number of successes (pn): Average waiting time = n pn = 1 p. (4.13) Note that the preceding reasoning gives us the average waiting time, without requiring any knowledge of the actual probability distribution of the waiting times (which we will calculate below). Of course, once we do know what the probability distribution is, we should be able to calculate the average (the expectation value) of the waiting times. This is the task of Problem 4.7. Distribution of waiting times Finding the probability distribution of the waiting times requires a little more work than finding the average waiting time. For the 10-sided die example, the question we re trying to answer is: What is the probability that if we consider two successive 1 s, the time between them will be 6 seconds? Or 30 seconds? Or 1 second? And so on. Although the average waiting time is 10 seconds, this certainly doesn t mean that the waiting time will always be 10 seconds. In fact, we will find below that the probability that the waiting time is exactly 10 seconds is quite small. Let s be general and say that the probability of success in each trial is p (so p = 1/10 in our present setup). Then the question is: What is the probability, P(k), that we will have to wait exactly k iterations (each of which is 1 second here) to obtain the next success? To answer this, note that in order for the next success to happen on the kth iteration, there must be failure (which happens with probability 1 p) on the first k 1 iterations, and then success on the kth one. The probability of this happening is P(k) = (1 p) k 1 p (geometric distribution) (4.14) This is the desired (discrete) probability distribution for the waiting time. This distribution goes by the name of the geometric distribution, because the probabilities form a geometric progression, due to the increasing power of the (1 p) factor. The geometric distribution is the discrete version of the exponential distribution that we ll arrive at in Eq. (4.26) below.

17 198 Chapter 4. Distributions Eq. (4.14) tells us that the probability that the next success comes on the very next iteration is p, the probability that it comes on the second iteration is (1 p)p, the probability that it comes on the third iteration is (1 p) 2 p, and so on. Each probability is smaller than the previous one by the factor (1 p). A plot of the distribution for p = 1/10 is shown in Fig The distribution is maximum at k = 1 and falls off from that value. Even though k = 10 is the average waiting time, the probability of the waiting time being exactly k = 10 is only P(10) = (0.9) 9 (0.1) 0.04 = 4% P(k) (p = 1/10) k Figure 4.12: The geometric distribution with p = 1/10. If p is large (close to 1), the plot of P(k) starts high (at p, which is close to 1) and then falls off quickly, because the factor (1 p) is close to 0. On the other hand, if p is small (close to 0), the plot of P(k) starts low (at p, which is close to 0) and then falls off slowly, because the factor (1 p) is close to 1. As a double check on the result in Eq. (4.14), we know that the next success has to eventually happen sometime, so the sum of all the P(k) probabilities must be 1. These P(k) probabilities form a geometric series whose first term is p and whose ratio is 1 p. The general formula for the sum of a geometric series with first term a and ratio r is a/(1 r), so we have P(1) + P(2) + P(3) + = p + p(1 p) + p(1 p) 2 + p = 1 (1 p) = 1, (4.15) as desired. As another check, we can verify that the expectation value (the average) of the waiting time for the geometric distribution in Eq. (4.14) equals 1/p, as we already found above; see Problem 4.7. You are encouraged to use a coin to experimentally verify Eq. (4.14) (or equivalently, the plot analogous to Fig. 4.12) for the case of p = 1/2. Just flip a coin as many times as you can in ten minutes, each time writing down a 1 if you get Heads and a 0 if you get Tails. Then make a long list of the waiting times between the 1 s. Then count up the number of one-toss waits, the number of two-toss waits, and so on. Then divide each of these numbers by the total number of waits (not the total number of tosses!) to find the probability of each waiting length. The results should

18 4.6. Exponential distribution 199 be (roughly) consistent with Eq. (4.14) for p = 1/2. In this case, the probabilities in Eq. (4.14) for k = 1, 2, 3, 4,... are 1/2, 1/4, 1/8, 1/16, Rates, expectation values, and probabilities Let s now consider the case where time is a continuous quantity. That is, let s assume that we can have a successful event at any instant, not just at the evenlyspaced 1-second marks as above. A continuous process whose probability is uniform in time can be completely described by just one number the average rate of success, which we ll call λ. We generally won t bother writing the word average, so we ll just call λ the rate. Before getting into the derivation of the continuous exponential distribution in Section 4.6.3, we ll need to talk a little about rates. The rate λ can be determined by counting the number of successful events that occur during a long time interval, and then dividing by this time. For example, if 300 (successful) events happen during 100 minutes, then the rate λ is 3 events per minute. Of course, if you count the number of events in a different span of 100 minutes, you will most likely get a slightly different number, perhaps 313 or 281. But in the limit of a very long time interval, you will find essentially the same rate, independent of which specific long interval you use. If the rate λ is 3 events per minute, you can alternatively write this as 1 event per 20 seconds, or 1/20 of an event per second. There is an infinite number of ways to write λ, and it s personal preference which one you pick. Just remember that you have to state the per time interval you re using. If you just say that the rate is 3, that doesn t mean anything. What is the expectation value of the number of events that happen during a time t? This expected number simply equals the product λt, from the definition of λ. If the expected number were anything other than λt, then if we divided it by t to obtain the rate, we wouldn t get λ. If you want to be a little more rigorous, consider a very large number n of intervals with length t. The total time in these intervals is nt. This total time is very large, so the number of events that happen during this time is (approximately) equal to (nt)λ, by the definition of λ. The expected number of events in each of the n intervals with length t is therefore ntλ/n = λt, as above. So we can write (Expected number of events in time t) = λt (4.16) In the above setup where λ equals 3 events per minute, the expected number of events that happen in, say, 5 minutes is λt = (3 events per minute)(5 minutes) = 15 events. (4.17) Does this mean that we are guaranteed to have exactly 15 events during a particular 5-minute span? Absolutely not. We can theoretically have any number of events, although there is essentially zero chance that the number will differ significantly from 15. (The probability of obtaining the various numbers of events is governed by the Poisson distribution, which we ll discuss in Section 4.7.) But the expectation value is 15. That is, if we perform a large number of 5-minute trials and then

19 200 Chapter 4. Distributions calculate the average number of events that occur in each trial, the result will be close to 15. A trickier question to ask is: What is the probability that exactly one event happens during a time t? Since λ is the rate, you might think that you can just multiply λ by t, as we did above, to say that the probability is λt. But this certainly can t be correct, because it would imply a probability of 15 for a 5-minute interval in the above setup. This is nonsense, because probabilities can t be larger than 1. If we instead pick a time interval of 20 seconds (1/3 of a minute), we obtain a λt value of 1. This doesn t have the fatal flaw of being larger than 1, but it has another issue, in that it says that exactly one event is guaranteed to happen during a 20-second interval. This can t be correct either, because it s certainly possible for zero (or two or three, etc.) events to occur. We ll figure out the exact probabilities of these numbers in Section 4.7. The strategy of multiplying λ by t to obtain a probability doesn t seem to work. However, there is one special case where it does work. If the time interval is extremely small (let s call it ϵ, which is a standard letter to use for something that is very small), then it is true that the probability of exactly one event occurring during the ϵ time interval is essentially equal to λϵ. We re using the word essentially because, although this statement is technically not true, it becomes arbitrarily close to being true in the limit where ϵ approaches zero. In the above example with λ = 1/20 events per second, the statement, λt is the probability that exactly one event happens during a time t, is a lousy approximation if t = 20 seconds, a decent approximation if t = 2 seconds, and a very good approximation if t = 0.2 seconds. And it only gets better as the time interval gets smaller. We ll explain why in the first remark below. We can therefore say that if P ϵ (1) stands for the probability that exactly one event happens during a small time interval ϵ, then P ϵ (1) λϵ (if ϵ is very small) (4.18) The smaller ϵ is, the better this approximation is. Technically, the condition in Eq. (4.18) is really if λϵ is very small. But we ll generally be dealing with normal sized λ s, so λϵ being small is equivalent to ϵ being small. When we deal with continuous time below, we ll actually be taking the ϵ 0 limit. In this mathematical limit, the sign in Eq. (4.18) becomes an exact = sign. To sum up: If t is very small, then λt is both the expected number of events that happen during the time t and (essentially) the probability that exactly one event happens during the time t. If t isn t very small, then λt is only the expected number of events. Remarks: 1. We claimed above that λt equals the probability of exactly one event occurring, only if t is very small. The reason for this restriction is that if t isn t small, then there is the possibility of multiple events occurring during the time t. We can be explicit about this as follows. Since we know from Eq. (4.16) that the expected number of events during

20 4.6. Exponential distribution 201 any time t is λt, we can use the expression for the expectation value in Eq. (3.4) to write λt = P t (0) 0 + P t (1) 1 + P t (2) 2 + P t (3) 3 +, (4.19) where P t (k) is the probability of obtaining exactly k events during the time t. Solving for P t (1) gives P t (1) = λt P t (2) 2 P t (3) 3 +. (4.20) We see that P t (1) is smaller than λt due to the P t (2) and P t (3), etc., probabilities. So P t (1) isn t equal to λt. However, if all of the probabilities of multiple events occurring (P t (2), P t (3), etc.) are very small, then P t (1) is essentially equal to λt. And this is exactly what happens if the time interval is very small. For small times, there is hardly any chance of the event even occurring once. So it is even less likely that it will occur twice, and even less likely for three times, etc. We can be a little more precise about this. The following argument isn t completely rigorous, but it should convince you that if t is very small, then P t (1) is essentially equal to λt. If t is very small, then assuming we don t know yet that P t (1) equals λt, we can still say that it should be roughly proportional to λt. This is true because if an event has only a tiny chance of occurring, then if you cut λ in half, the probability is essentially cut in half. Likewise if you cut t in half. This proportionality then implies that the probability that exactly two events occur is essentially proportional to (λt) 2. We ll see in Section 4.7 that there is actually a factor of 1/2 involved here, but that is irrelevant in the present argument. The important point is the quadratic nature of (λt) 2. If λt is sufficiently small, then (λt) 2 is negligible compared with λt. Likewise for P t (3) (λt) 3, etc. We can therefore ignore the scenarios where multiple events occur. So with t ϵ, Eq. (4.20) becomes P ϵ (1) λϵ P ϵ (2) 2 P ϵ (3) 3 +, (4.21) in agreement with Eq. (4.18). As mentioned above, if λϵ is small, it is because ϵ is small, at least in the situations we ll be dealing with. 2. Imagine drawing the λ vs. t curve. We have put curve in quotes because the curve is actually just a straight horizontal line, since we re assuming a constant λ. If we consider a time interval t, the associated area under the curve equals λ t, because we have a simple rectangular region. So from Eq. (4.18), this area gives the probability that an event occurs during a time t, provided that t is very small. This might make you think that λ can be interpreted as a probability distribution, because we found in Section that the area under a distribution curve gives the probability. However, the λ curve cannot be interpreted as a probability distribution, because this areaequals-probability result holds only for very small t. The area under a distribution curve has to give the probability for any interval on the horizontal axis. The λ curve doesn t satisfy this property. The total area under the λ curve is infinite (because the straight horizontal line extends for all time), whereas actual probability distributions must have a total area of Since only one quantity, λ, is needed to describe everything about a random process whose probability is uniform in time, any other quantity we might want to determine must be able to be written in terms of λ. This will become evident below.